Skip to content

Latest commit

 

History

History
333 lines (240 loc) · 13.1 KB

README.md

File metadata and controls

333 lines (240 loc) · 13.1 KB
Raw

Librsvg test suite

Librsvg's test suite is split like this:

  • Unit tests in the Rust code, run normally with "cargo test".

  • Rust integration tests in this tests/ directory.

  • C API tests in this tests/ directory.

The C API and Rust tests run the library with its public APIs in both languages. In addition, the Rust tests also exercise the rsvg-convert program.

For the impatient: you can use cargo test to run most of the test suite; this is fine for regular development. This will not run the C API tests or some of the long-running tests that exercise the hard-coded limits of the library.

To run the full test suite, see "Running the test suite" below.

Unit tests

The library's source code has small unit tests for particular sections of the code.

It is better to catch errors early, in the unit tests, if possible. The test suite in this tests/ directory is for black box tests, which run the library as a normal program would use it.

  • What should be in a unit test - a small test of an algorithm; a check for computed values given some starting values; checks for edge cases.

  • What should be in these black-box tests - rendering tests that exercise a particular part of the code; CSS cascading tests; images that expose bugs and that we want to avoid regressing on later.

For example, there are unit tests of the path data parser (the <path d="M10 10 L20 20 ..."> element and its d attribute, to ensure that the parser handles all the path commands and catches errors appropriately. Correspondingly, there are a bunch of black-box tests that exercise particular features of path rendering ("does this actually draw a line, or an arc?").

Running the test suite

For regular development, use cargo test. This will run most of the test suite, except for the C API tests and the long-running tests which exercise the hard-coded limits of the library.

To run the full test suite, you need to go through autotools. Run the following commands in the toplevel source directory:

./autogen.sh
make check

C API tests - api.c

These test the full C API of librsvg: all the public functions; the RsvgHandle class, its methods, and its GObject properties; all the deprecated functions. Any new public APIs should get tested here.

The tests are intended to preserve the historical peculiarities of the C API, to ensure ABI compatibility across versions of the library.

These tests are not meant to exhaustively test librsvg's features. For those, you should look at the [Rust integration tests][#rust-integration-tests].

This C API test suite is built upon Glib's GTest utility functions, which let you define tests in the C language.

Rust integration tests

These are built as a Rust binary in this tests/ directory, and are runnable with cargo test.

Rust API tests - api.rs

Tests the public Rust API of librsvg.

Crash tests - loading_crash.rs

These load and parse an SVG, and ensure that there are no crashes in the process. Note that this does not render the images.

The SVG images live in the fixtures/crash directory. The files are just tested to not cause crashes during the loading process; it does not matter if the files are well-formed XML, for example.

Rendering crash tests - render_crash.rs

We use these tests to ensure there are no regressions after fixing a bug where a particular SVG loads fine, but it crashes the renderer.

The test files are in the fixtures/render-crash directory. The module loads the files and renders them, without comparing the results to anything in particular.

General bug regression tests - bugs.rs

These test fixes for specific bugs in the library, so that the bugs don't recur.

Error tests - errors.rs

These test conditions which should produce errors during loading or rendering.

During loading, librsvg will report malformed XML as errors. It will also report an error if an SVG file has more elements than what is configured in librsvg's internal limits; this is intended to prevent uncontrolled memory consumption.

During rendering, librsvg will report errors if the SVG document hits librsvg's internal limits for the number of instanced objects; this is intended to prevent uncontrolled CPU consumption from malicious SVG files.

The test files are in the fixtures/errors directory.

Tests for SVG filter effects - filters.rs

These test the semantics of the filter property, and specific filter functions.

Reference tests - reference.rs

These are the bulk of the rendering tests, where the results of rendering SVG files are compared against reference PNG images.

The reference tests allow for minor differences in the pixel values of the results. Each pixel's RGBA components gets compared to the corresponding one in the reference image:

  • If the absolute value of the difference between corresponding RGBA components is more than 2, the test suite considers the result to be distinguishable from the reference, but otherwise acceptable.

  • If the absolute value of the difference is more than the number in the RSVG_TEST_TOLERANCE environment variable, the result is inacceptable and the test suite fails; the default is 2 if that variable is not set. You can tweak this value if your machine's floating-point unit produces wildly different results.

The test files are in the fixtures/reftests/ directory. Each image-based reference test uses two files: foo.svg and foo-ref.png. The test harness will render foo.svg and compare the results to foo-ref.png.

Failing tests will appear as part of the cargo test output. It will print the filenames for the output and difference images for failed tests, as follows.

Each foo.svg test file produces a foo-out.png result, and if that result is distinguishable from the reference PNG (per the terminology above), the test will also produce a foo-diff.png which you can examine by hand. See "Examining failed reference tests" below.

Ignoring tests: SVG test files in fixtures/reftests whose names begin with "ignore" will be skipped from the tests. That is, anything that matches "ignore*.svg`" will not be included in the tests. You can use this to skip a few problematic files temporarily.

As of 2020/Oct/22 we have an informal organization of these files:

  • fixtures/reftests/svg1.1 - Tests from the W3C's SVG1.1 test suite. These are supposed to test all of SVG's features; we will add them one by one as librsvg starts implementing the features.

  • fixtures/reftests/svg2 - Tests for SVG2 or CSS3 features.

  • fixtures/reftests/bugs/*.svg - Tests for particular bug numbers. Please use the bug number from Gitlab, like 1234-blah.svg, and the corresponding 1234-blah-ref.png for the known-good reference image.

    Note: Librsvg migrated from git.gnome.org and bugzilla.gnome.org to gitlab.gnome.org. Bug numbers in Bugzilla were around 6 digits in length; in Gitlab, they are small numbers.

  • fixtures/reftests/*.svg - Tests for special situations that arose during development.

  • fixtures/reftests/adwaita/*.svg - A snapshot of the Adwaita icon theme (GNOME's default icon theme), to ensure that librsvg renders it correctly.

Examining failed reference tests

Let's say you run make check and see that one of the tests fails. The test log may have lines like these:

---- reference::svg_1_1_tests_fixtures_reftests_svg1_1_painting_stroke_01_t_svg stdout ----
output: output/painting-stroke-01-t-out.png
painting-stroke-01-t: 12414 pixels changed with maximum difference of 255
diff: output/painting-stroke-01-t-diff.png
thread 'reference::svg_1_1_tests_fixtures_reftests_svg1_1_painting_stroke_01_t_svg' panicked at 'surfaces are too different', tests/src/reference.rs:319:25

This means that the test file fixtures/reftests/svg1.1/painting-stroke-01-t.svg got rendered, and produced incorrect output when compared to fixtures/reftests/svg1.1/painting-stroke-01-t-ref.png.

When a test fails, rsvg-test creates two images in tests/output:

tests/output/foo-out.png
tests/output/foo-diff.png

In this case, foo-out.png is the actual rendered output, which is presumed to be incorrect, since it differs from the foo-ref.png reference image. The foo-diff.png is a "visual diff" that you can see in an image viewer; pixels that differ are highlighted.

It is up to you to decide what to do next:

  • If the foo-out.png image looks correct, and the only difference with respect to the foo-ref.png reference image is that antialiased edges look different, or font rendering is slightly different due to the font-rendering machinery in your system, you can just regenerate the test image. See "Regenerating reference images" below.

  • If the foo-out.png image is obviously wrong when compared to the foo-ref.png reference, you can file a bug. You can wait until someone fixes it, or try to fix the bug yourself!

  • Any other situation of course deserves attention. Feel free to ask the maintainers about it; even if you figure out the problem yourself, a failed test almost always indicates a problem that is not just on your end.

Regenerating reference images

Let's say the test tests/fixtures/reftests/.../foo.svg failed. Then you fix the bug, or determine that the output image is in fact correct, and it just differs from the reference image due to antialiasing artifacts. In this case, your next step is to regenerate the reference image so the test passes again.

You should not just use rsvg-convert to render test files! The test machinery sets up conditions for [reproducible font rendering][#reproducible-font-rendering], which are not available to rsvg-convert.

Run cargo test, and copy the resulting foo-out.png to the tests/fixtures/.../foo-ref.png that corresponds to foo.svg.

You can then run cargo test again and ensure that the tests pass.

Issues with the official SVG test suite

Our SVG files in tests/fixtures/reftests/svg1.1 come from the "SVG 1.1 Second Edition test suite" archive linked here:

https://www.w3.org/Graphics/SVG/WG/wiki/Test_Suite_Overview

We don't know how the reference PNG files in that archive are generated. However, they are done in such a way that objects tend not to be pixel-aligned. For example, many tests have a rectangular frame around the whole viewport, defined like this:

<rect id="test-frame" x="1" y="1" width="478" height="358" fill="none" stroke="#000000"/>

This specifies no stroke width, so it uses 1 by default. The desired effect is "stroke this rectangle with a 1-pixel wide line".

However, notice that the (x, y) coordinates of the rect are (1, 1). This means that the actual bounds of the stroked outline are from (0.5, 0.5) to (479.5, 359.5). The result is a fuzzy outline: it occupies two pixels of width, with each pixel having half-black coverage.

Some elements in the reference PNG images from the official SVG test suite are pixel-aligned, and some are not, like the example test-frame above. It looks like their renderer uses a (0.5, 0.5) offset just for strokes, but not for fills, which sounds hackish.

Our test suite does not use special offsets, so that SVG images not from the official test suite are rendered "normally". This means that the reference images from the official test suite will always fail initially, since stroke outlines will be fuzzy in librsvg, but not in the test suite (and conversely, SVGs not from the test suite would be crisp in librsvg but probably not in the test suite's renderer renderer).

Also, the original reference PNGs from the SVG 1.1 test suite either use fonts that are different from those usually on free software systems, or they use SVG fonts which librsvg currently doesn't support (i.e. with glyph shapes referenced from a secondary SVG).

In any case, look at the results by hand, and compare them by eye to the official reference image. If the thing being tested looks correct, and just the outlines are fuzzy — and also it is just the actual font shapes that are different — then the test is probably correct. Follow the procedure as in "Regenerating reference images" listed above in order to have a reference image suitable for librsvg.

Reproducible font rendering

The test runners set up special conditions so that font rendering is reproducible across systems. Normally, font rendering can vary quite a bit depending on various factors:

  • Versions of fontconfig, freetype, cairo, and pango.
  • Installed fonts.
  • The system's font mappings.
  • The user's settings for font antialiasing, hinting, etc.

The test suite includes part of the Roboto fonts in librsvg/tests/resources, and creates a configuration font map with just those fonts. In addition, the Pango context used for rendering is set up with a hardcoded mode for antialiasing, hinting, and hint metrics.